Two Real Product Decisions
What distinguishes Muse Code isn't that it writes code — it's two engineering choices. First, subagents persist: a main agent coordinates background workers running in parallel inside isolated worktrees, and those workers stay active for the whole session rather than being spawned per task, which avoids repeatedly re-gathering the same context. Second, the runtime is replayable: a local event log appends every model call, tool execution, approval, and file edit, which Meta describes as "replay-exact and restart-safe" — after a crash the agent resumes exactly where it stopped, which is what long migrations and large refactors actually need. Three skills ship built in: `/plan` turns a task into an approval-gated plan, `/grill` stress-tests that plan, and `/goal` drives toward a stated objective.
The Model and the Scores
Muse Spark 1.2 is a coding-focused checkpoint co-trained with Muse Code, with a 1M-token context. Meta's own numbers put it at 82.9% on Terminal-Bench 2.1 (all 89 tasks, pass@1 over five attempts) and 59.3% on DeepSWE v1.1 (113 tasks across 91 repositories and five languages), against 76.2% and 53.0% for the previous Muse Spark 1.1. Worth stating plainly: all three evaluations were run by Meta, and its own methodology document concedes the setup may not be tuned to third-party models' strengths. The Register notes the scores are tightly clustered across vendors and that Muse Spark did not come out on top.
The Play Is Price, Not Capability
Meta AI head Alexandr Wang was direct about it: this release competes with Anthropic's and OpenAI's products on price, not capability. The default pay-as-you-go tier is $1.25 per million input tokens and $4.25 per million output; a cheaper "contributor" tier is available to developers who agree to have their data used for training. Both are served through the Meta Model API and can also be reached via OpenRouter. One more thing worth noting: Muse Spark is a closed-weight, cloud-hosted model — a departure from the open-weight approach of the Llama era — and Zuckerberg would only say he'd "have more to share on that soon." For developers, what's certain today is one more terminal agent with a lower price and a longer context. Whether it displaces the tool you already use depends on your codebase size and how the subagent machinery holds up in a real project, not on these benchmark numbers.
via: Meta AI Research: Introducing Muse Code and Muse Spark 1.2, The Register report, TechCrunch: Meta launches Muse Code, an AI agent for large code bases; verified 2026-08-07